Deep learning methods for forecasting of extreme ambient ozone values

Exposure to high ozone concentrations can be harmful for humans and therefore many countries have declared a threshold for ozone concentrations that should not be exceeded. That is why the ability to predict high and extreme near surface ozone concentrations is important not only for human health but also for regulatory purposes. The problem with many existing ozone forecasting methods, especially deep learning approaches, is their inaccuracy and unreliability to forecast high ozone concentrations. The goal of this study is to discover the usage of oversampling and subsequent finetuning to increase the forecast precision for extreme near surface ozone concentration. Therefore the architecture and experiment setup of IntelliO3-ts, a convolutional neural network for the forecast of near surface ozone concentrations, is used as a foundation to which the methods are applied. At first, oversampling is applied to the data set, which is the process of multiplying samples from less frequent ozone concentration ranges and adding them to the data set. The thereby obtained new "oversampled" data set that has a flatter sample distribution is then used to train the neural network. In a second and additional step the finetuning takes place, which is a retraining of the network obtained in the first step, using the original data set before oversampling was applied. For both methods different parameters will be tested and evaluated on the basis of different scores calculated on 2x2 contingency tables. The contingency tables are created by using a threshold and separating the test data in two groups, ozone concentrations below and above the threshold. The oversampling increases the ability to successfully forecast if a sample exceeds a certain threshold, thereby increasing the forecast precision for high ozone values. These advantages come at the cost of also increasing the percentage of samples that are falsely predicted to be above a certain threshold, also resulting in a systematic overestimation. The best model obtained, was able to increase the hit rate at 60 ppb from 43% to 67% and at 80 ppb from 1:9% to 15.2%. This means that the modelis able to correctly predict that a sample is above 60 ppb and 80 ppb for 67% and 15.2% of all samples above that threshold instead of only achieving this for 43% and1.9% of the samples above that threshold, respectively. Therefore the oversampling offers a valuable trade off, to sacrifice parts of the overall performance in order to increase the ability to forecast high ozone values, which might be useful especially for regulatory purposes. The finetuning did not add any new value to that, but only reverted some of the improvements that were achieved by the oversampling. For future research it might be interesting to investigate the usage of different oversampling methods or explore the application of oversampling and finetuning to the forecasting of multiple days, as this study only focused on the forecast for the next day